Context-Aware Asymmetric Ensembling for Interpretable Retinopathy of Prematurity Screening via Active Query and Vascular Attention

TL;DR

This work tackles automated retinopathy of prematurity screening under data scarcity and class imbalance by introducing a Context-Aware Asymmetric Ensemble (CAA Ensemble) that separates structural and vascular analysis into two specialized streams. The structure-focused MS-AQNet is guided by clinical priors via an active-query mechanism, while the texture-focused VascuMIL leverages vascular topology maps within a MIL framework, and a fusion meta-learner combines their signals with metadata for robust, explainable decisions. Ablation studies show that inductive bias from asymmetric ensembling, VMAP-guided texture analysis, and metadata-conditioned attention collectively yield state-of-the-art results on a small public cohort, with strong interpretability through heatmaps and vascular threat maps. The approach reduces dependence on large private datasets and offers a practical, transparent framework for clinical deployment and telemedicine triage in diverse settings.

Abstract

Retinopathy of Prematurity (ROP) is among the major causes of preventable childhood blindness. Automated screening remains challenging, primarily due to limited data availability and the complex condition involving both structural staging and microvascular abnormalities. Current deep learning models depend heavily on large private datasets and passive multimodal fusion, which commonly fail to generalize on small, imbalanced public cohorts. We thus propose the Context-Aware Asymmetric Ensemble Model (CAA Ensemble) that simulates clinical reasoning through two specialized streams. First, the Multi-Scale Active Query Network (MS-AQNet) serves as a structure specialist, utilizing clinical contexts as dynamic query vectors to spatially control visual feature extraction for localization of the fibrovascular ridge. Secondly, VascuMIL encodes Vascular Topology Maps (VMAP) within a gated Multiple Instance Learning (MIL) network to precisely identify vascular tortuosity. A synergistic meta-learner ensembles these orthogonal signals to resolve diagnostic discordance across multiple objectives. Tested on a highly imbalanced cohort of 188 infants (6,004 images), the framework attained State-of-the-Art performance on two distinct clinical tasks: achieving a Macro F1-Score of 0.93 for Broad ROP staging and an AUC of 0.996 for Plus Disease detection. Crucially, the system features `Glass Box' transparency through counterfactual attention heatmaps and vascular threat maps, proving that clinical metadata dictates the model's visual search. Additionally, this study demonstrates that architectural inductive bias can serve as an effective bridge for the medical AI data gap.

Figures (6)

• Figure 1: Overview of the Context-Aware Asymmetric Ensemble. (A) Intelligent Data Engineering: Morphological artifact removal and geometry-preserving resizing ($384 \times 384$) for structural analysis. (B) Vascular Topology Extraction: Generation of VMAPs from high-resolution inputs ($768 \times 768$) to form 4-channel tensors for textural analysis. (C) Processing & Fusion: Parallel execution of the Structure Specialist (MS-AQNet) and Texture Specialist (VascuMIL), followed by synergistic fusion of logits with re-injected clinical metadata.
• Figure 2: Architecture of the Multi-Scale Active Query Network (MS-AQNet). (A) The hierarchical pipeline extracting features at three spatial scales via an EfficientNet-B0 backbone. (B) Detail of the Active Query Unit, showing how the projected clinical query ($\mathbf{q}_s$) generates a spatial attention map via dot-product similarity, regulated by a learnable gate ($\alpha$). (C) Detail of the FiLM Block, illustrating the global affine transformation ($\gamma, \beta$) of the visual embedding based on clinical priors.
• Figure 3: Architecture of the VascuMIL Network.(Top)Feature Extraction: 4-channel input tensors (RGB + Vascular Topology Maps) are encoded via a weight-shared backbone with a modified stem. (Bottom)Gated Attention: A learnable gating mechanism ($\tanh \odot \sigma$) assigns importance weights to instances, isolating pathological signals (Red) from background noise (Grey) to aggregate the final patient context vector $\mathbf{z}$.
• Figure 4: Diagnostic Performance Analysis. (A-B) Normalized confusion matrices for Broad Diagnosis and Plus Disease, demonstrating high diagonal dominance and minimal critical errors. (C) Per-class ROC curves for the Ensemble model, showing robust separation across all diagnostic categories ($AUC > 0.99$). (D) ROC curves for Plus Disease detection, highlighting the superior AUC of the Ensemble (Red) compared to individual structure (Blue) and texture (Green) specialists.
• Figure 5: Multi-Modal Clinical Explainability. Comparison of diagnostic attention maps. (Middle) MS-AQNet localizes global structural anomalies (e.g., ridges in Row 2). (Right) VascuMIL isolates micro-vascular tortuosity (Red/Yellow) from background (Blue).